@kkelly-offical/kkcode 0.1.3 → 0.1.7

package/src/session/compaction.mjs

@@ -1,276 +1,298 @@
-import { requestProvider } from "../provider/router.mjs"
-import { getConversationHistory, replaceMessages } from "./store.mjs"
-import { HookBus } from "../plugin/hook-bus.mjs"
-import { saveCheckpoint } from "./checkpoint.mjs"
-import { recordTurn } from "../usage/usage-meter.mjs"
-import { loadPricing, calculateCost } from "../usage/pricing.mjs"
-
-const COMPACTION_SYSTEM = `You are a conversation summarizer. Create a structured summary preserving all critical information for continued work.
-
-## Output Format
-
-<summary>
-<goal>The user's overall goal or current task</goal>
-<completed>
-- Completed task with specific details (file paths, function names, line numbers)
-</completed>
-<in_progress>Current work being done, if any</in_progress>
-<files_modified>
-- path/to/file: specific change description
-</files_modified>
-<key_decisions>
-- Decision and reasoning
-- User preferences or constraints
-</key_decisions>
-<errors_resolved>
-- Error description → fix applied
-</errors_resolved>
-<next_steps>
-- Specific next action items
-</next_steps>
-</summary>
-
-Rules:
-- Use the SAME LANGUAGE as the conversation
-- Preserve ALL file paths, function names, variable names, and technical identifiers exactly
-- Include specific code changes, not just "modified file X"
-- Omit tool call metadata and message formatting details
-- Be concise but never drop actionable information`
-
-const DEFAULT_THRESHOLD_MESSAGES = 50
-const DEFAULT_THRESHOLD_RATIO = 0.7
-const DEFAULT_KEEP_RECENT = 6
-const TOOL_RESULT_PREVIEW_LIMIT = 200
-
-// Estimate tokens from a string, accounting for CJK characters (~1.5 chars/token vs ~4 for Latin)
-export function estimateStringTokens(str) {
-  if (!str) return 0
-  let cjk = 0
-  for (let i = 0; i < str.length; i++) {
-    const code = str.charCodeAt(i)
-    if ((code >= 0x4E00 && code <= 0x9FFF) || (code >= 0x3000 && code <= 0x30FF) ||
-        (code >= 0xAC00 && code <= 0xD7AF)) cjk++
-  }
-  const latin = str.length - cjk
-  return Math.ceil(latin / 4 + cjk / 1.5)
-}
-
-const MSG_OVERHEAD = 4 // ~4 tokens per message for role/metadata
-
-export function estimateTokenCount(messages) {
-  let tokens = 0
-  for (const msg of messages) {
-    tokens += MSG_OVERHEAD
-    const content = msg.content
-    if (Array.isArray(content)) {
-      for (const block of content) {
-        if (block.type === "image") {
-          tokens += 1600 // conservative estimate for a typical image
-        } else if (block.type === "tool_use") {
-          tokens += estimateStringTokens(block.name || "")
-          tokens += estimateStringTokens(JSON.stringify(block.input || {}))
-        } else if (block.type === "tool_result") {
-          tokens += estimateStringTokens(String(block.content || ""))
-        } else {
-          tokens += estimateStringTokens(block.text || block.content || "")
-        }
-      }
-    } else {
-      tokens += estimateStringTokens(content || "")
-    }
-  }
-  return tokens
-}
-
-/**
- * Pre-prune messages before LLM summarization.
- * - Truncate large tool_result content to a short preview
- * - Keep tool_use blocks intact (they show model intent)
- * - Truncate very long plain-text assistant/user messages
- */
-export function pruneForSummary(messages, previewLimit = TOOL_RESULT_PREVIEW_LIMIT) {
-  return messages.map((msg) => {
-    const content = msg.content
-    if (Array.isArray(content)) {
-      const pruned = content.map((block) => {
-        if (block.type === "tool_result") {
-          const raw = String(block.content || "")
-          if (raw.length > previewLimit) {
-            return {
-              ...block,
-              content: `${raw.slice(0, previewLimit)}... [truncated ${raw.length} chars]`
-            }
-          }
-        }
-        return block
-      })
-      return { ...msg, content: pruned }
-    }
-    // Truncate very long plain-text messages (e.g. large tool output pasted as text)
-    if (typeof content === "string" && content.length > 2000) {
-      return { ...msg, content: `${content.slice(0, 2000)}... [truncated ${content.length} chars]` }
-    }
-    return msg
-  })
-}
-
-const BUILTIN_CONTEXT = {
-  "gpt-5": 272000, "o3": 200000, "o1": 200000,
-  "claude-opus-4": 200000, "claude-3-5": 200000, "claude-3.5": 200000, "claude": 200000,
-  "gemini-2": 1048576, "gemini-1.5": 1048576, "gemini": 128000,
-  "gpt-4o": 128000, "gpt-4": 128000, "gpt-3.5": 16000,
-  "deepseek": 64000, "qwen": 128000
-}
-
-export function modelContextLimit(model, configState = null) {
-  const m = String(model || "").toLowerCase()
-  // 1) Check provider-level context_limit for the active provider
-  const providerCfg = configState?.config?.provider
-  if (providerCfg) {
-    // Per-model override from provider.model_context map
-    const mc = providerCfg.model_context
-    if (mc) {
-      if (mc[model]) return mc[model]
-      for (const key of Object.keys(mc)) {
-        if (m.startsWith(key.toLowerCase())) return mc[key]
-      }
-    }
-    // Provider-level context_limit
-    const active = providerCfg[providerCfg.default]
-    if (active?.context_limit > 0) return active.context_limit
-  }
-  // 2) Builtin prefix match
-  for (const [prefix, limit] of Object.entries(BUILTIN_CONTEXT)) {
-    if (m.includes(prefix)) return limit
-  }
-  return 128000
-}
-
-export function contextUtilization(messages, model, configState = null) {
-  const tokens = estimateTokenCount(messages)
-  const limit = modelContextLimit(model, configState)
-  const ratio = limit > 0 ? Math.min(1, tokens / limit) : 0
-  return {
-    tokens,
-    limit,
-    ratio,
-    percent: Math.round(ratio * 100)
-  }
-}
-
-export function supportsNativeCompaction(providerType, model) {
-  if (providerType !== "anthropic") return false
-  const m = String(model || "").toLowerCase()
-  return m.includes("claude") && (m.includes("opus") || m.includes("sonnet"))
-}
-
-export function shouldCompact({ messages, model, thresholdMessages = DEFAULT_THRESHOLD_MESSAGES, thresholdRatio = DEFAULT_THRESHOLD_RATIO, configState = null, realTokenCount = null }) {
-  if (messages.length >= thresholdMessages) return true
-  const limit = modelContextLimit(model, configState)
-  const tokens = realTokenCount != null ? realTokenCount : estimateTokenCount(messages)
-  return tokens >= limit * thresholdRatio
-}
-
-export async function compactSession({
-  sessionId,
-  model,
-  providerType,
-  configState,
-  keepRecent = DEFAULT_KEEP_RECENT,
-  baseUrl = null,
-  apiKeyEnv = null
-}) {
-  const history = await getConversationHistory(sessionId, 9999)
-  if (history.length <= keepRecent + 2) return { compacted: false, reason: "too few messages" }
-
-  // Find split point that doesn't break tool_use/tool_result pairs
-  let splitIdx = history.length - keepRecent
-  while (splitIdx > 0 && splitIdx < history.length) {
-    const msg = history[splitIdx]
-    const content = msg.content
-    if (Array.isArray(content) && content.some(b => b.type === "tool_result")) {
-      splitIdx-- // include the paired assistant tool_use message
-      continue
-    }
-    break
-  }
-  const toSummarize = history.slice(0, splitIdx)
-  const kept = history.slice(splitIdx)
-
-  // Layer 1: prune large tool outputs before sending to LLM
-  const pruned = pruneForSummary(toSummarize)
-  const summaryPrompt = pruned.map((m) => {
-    const content = m.content
-    if (Array.isArray(content)) {
-      return `[${m.role}]: ${content.map((b) => {
-        if (b.type === "text") return b.text || ""
-        if (b.type === "tool_use") return `[tool_use:${b.name}(${JSON.stringify(b.input || {}).slice(0, 120)})]`
-        if (b.type === "tool_result") return `[tool_result:${b.is_error ? "ERROR " : ""}${b.content || ""}]`
-        return ""
-      }).filter(Boolean).join("\n")}`
-    }
-    return `[${m.role}]: ${content}`
-  }).join("\n\n")
-
-  const hookPayload = await HookBus.sessionCompacting({
-    sessionId,
-    messageCount: history.length,
-    summarizeCount: toSummarize.length,
-    keepCount: kept.length
-  })
-  if (hookPayload?.skip) return { compacted: false, reason: "skipped by hook" }
-
-  let summaryText
-  let compactionUsage = null
-  try {
-    const response = await requestProvider({
-      configState,
-      providerType,
-      model,
-      system: COMPACTION_SYSTEM,
-      messages: [{ role: "user", content: summaryPrompt }],
-      tools: [],
-      baseUrl,
-      apiKeyEnv
-    })
-    summaryText = (response.text || "").trim()
-    compactionUsage = response.usage || null
-  } catch (error) {
-    return { compacted: false, reason: `compaction LLM call failed: ${error.message}` }
-  }
-
-  if (!summaryText) return { compacted: false, reason: "empty summary from LLM" }
-
-  // Replace all messages with: [summary] + [kept recent messages]
-  const summaryMessage = {
-    role: "user",
-    content: `<compaction-summary>\n${summaryText}\n</compaction-summary>`
-  }
-  await replaceMessages(sessionId, [summaryMessage, ...kept])
-
-  // Record compaction LLM usage so it's not "invisible"
-  if (compactionUsage) {
-    try {
-      const { pricing } = await loadPricing(configState)
-      const { amount } = calculateCost(pricing, model, compactionUsage)
-      await recordTurn({ sessionId, usage: compactionUsage, cost: amount })
-    } catch { /* best-effort */ }
-  }
-
-  await saveCheckpoint(sessionId, {
-    kind: "compaction",
-    iteration: 0,
-    compactedAt: Date.now(),
-    summarizeCount: toSummarize.length,
-    keepCount: kept.length,
-    summaryVersion: 1,
-    summaryLength: summaryText.length
-  })
-
-  return {
-    compacted: true,
-    summarizedCount: toSummarize.length,
-    keptCount: kept.length,
-    summaryLength: summaryText.length
-  }
-}
+import { requestProvider } from "../provider/router.mjs"
+import { getConversationHistory, replaceMessages } from "./store.mjs"
+import { HookBus } from "../plugin/hook-bus.mjs"
+import { saveCheckpoint } from "./checkpoint.mjs"
+import { recordTurn } from "../usage/usage-meter.mjs"
+import { loadPricing, calculateCost } from "../usage/pricing.mjs"
+
+const COMPACTION_SYSTEM = `You are a conversation summarizer. Create a structured summary preserving all critical information for continued work.
+
+## Output Format
+
+<summary>
+<goal>The user's overall goal or current task</goal>
+<completed>
+- Completed task with specific details (file paths, function names, line numbers)
+</completed>
+<in_progress>Current work being done, if any</in_progress>
+<files_modified>
+- path/to/file: specific change description
+</files_modified>
+<key_decisions>
+- Decision and reasoning
+- User preferences or constraints
+</key_decisions>
+<errors_resolved>
+- Error description → fix applied
+</errors_resolved>
+<next_steps>
+- Specific next action items
+</next_steps>
+</summary>
+
+Rules:
+- Use the SAME LANGUAGE as the conversation
+- Preserve ALL file paths, function names, variable names, and technical identifiers exactly
+- Include specific code changes, not just "modified file X"
+- Omit tool call metadata and message formatting details
+- Be concise but never drop actionable information`
+
+const DEFAULT_THRESHOLD_MESSAGES = 50
+const DEFAULT_THRESHOLD_RATIO = 0.7
+const DEFAULT_KEEP_RECENT = 6
+const DEFAULT_KEEP_RECENT_TURNS = 3
+const TOOL_RESULT_PREVIEW_LIMIT = 200
+
+// Estimate tokens from a string, accounting for CJK characters (~1.5 chars/token vs ~4 for Latin)
+export function estimateStringTokens(str) {
+  if (!str) return 0
+  let cjk = 0
+  for (let i = 0; i < str.length; i++) {
+    const code = str.charCodeAt(i)
+    if ((code >= 0x4E00 && code <= 0x9FFF) || (code >= 0x3000 && code <= 0x30FF) ||
+        (code >= 0xAC00 && code <= 0xD7AF)) cjk++
+  }
+  const latin = str.length - cjk
+  return Math.ceil(latin / 4 + cjk / 1.5)
+}
+
+const MSG_OVERHEAD = 4 // ~4 tokens per message for role/metadata
+
+export function estimateTokenCount(messages) {
+  let tokens = 0
+  for (const msg of messages) {
+    tokens += MSG_OVERHEAD
+    const content = msg.content
+    if (Array.isArray(content)) {
+      for (const block of content) {
+        if (block.type === "image") {
+          tokens += 1600 // conservative estimate for a typical image
+        } else if (block.type === "tool_use") {
+          tokens += estimateStringTokens(block.name || "")
+          tokens += estimateStringTokens(JSON.stringify(block.input || {}))
+        } else if (block.type === "tool_result") {
+          tokens += estimateStringTokens(String(block.content || ""))
+        } else {
+          tokens += estimateStringTokens(block.text || block.content || "")
+        }
+      }
+    } else {
+      tokens += estimateStringTokens(content || "")
+    }
+  }
+  return tokens
+}
+
+/**
+ * Pre-prune messages before LLM summarization.
+ * - Strip synthetic scaffolding messages (continuation noise)
+ * - Truncate large tool_result content with aging: older steps get shorter previews
+ * - Keep tool_use blocks intact (they show model intent)
+ * - Truncate very long plain-text assistant/user messages
+ */
+export function pruneForSummary(messages, previewLimit = TOOL_RESULT_PREVIEW_LIMIT) {
+  // Strip synthetic scaffolding messages (continuation prompts, fake tool_result errors)
+  const real = messages.filter(msg => !msg.synthetic)
+
+  // #2 工具结果老化: find max step to compute relative age per message
+  const maxStep = real.reduce((m, msg) => Math.max(m, msg.step || 0), 0)
+
+  return real.map((msg) => {
+    // Aging: older tool_results get more aggressive truncation
+    const age = maxStep - (msg.step || 0)
+    const effectiveLimit = Math.max(50, previewLimit - age * 15)
+
+    const content = msg.content
+    if (Array.isArray(content)) {
+      const pruned = content.map((block) => {
+        if (block.type === "tool_result") {
+          const raw = String(block.content || "")
+          if (raw.length > effectiveLimit) {
+            return {
+              ...block,
+              content: `${raw.slice(0, effectiveLimit)}... [truncated ${raw.length} chars, age=${age}]`
+            }
+          }
+        }
+        return block
+      })
+      return { ...msg, content: pruned }
+    }
+    // Truncate very long plain-text messages (e.g. large tool output pasted as text)
+    if (typeof content === "string" && content.length > 2000) {
+      return { ...msg, content: `${content.slice(0, 2000)}... [truncated ${content.length} chars]` }
+    }
+    return msg
+  })
+}
+
+const BUILTIN_CONTEXT = {
+  "gpt-5": 272000, "o3": 200000, "o1": 200000,
+  "claude-opus-4": 200000, "claude-3-5": 200000, "claude-3.5": 200000, "claude": 200000,
+  "gemini-2": 1048576, "gemini-1.5": 1048576, "gemini": 128000,
+  "gpt-4o": 128000, "gpt-4": 128000, "gpt-3.5": 16000,
+  "deepseek": 64000, "qwen": 128000
+}
+
+export function modelContextLimit(model, configState = null) {
+  const m = String(model || "").toLowerCase()
+  // 1) Check provider-level context_limit for the active provider
+  const providerCfg = configState?.config?.provider
+  if (providerCfg) {
+    // Per-model override from provider.model_context map
+    const mc = providerCfg.model_context
+    if (mc) {
+      if (mc[model]) return mc[model]
+      for (const key of Object.keys(mc)) {
+        if (m.startsWith(key.toLowerCase())) return mc[key]
+      }
+    }
+    // Provider-level context_limit
+    const active = providerCfg[providerCfg.default]
+    if (active?.context_limit > 0) return active.context_limit
+  }
+  // 2) Builtin prefix match
+  for (const [prefix, limit] of Object.entries(BUILTIN_CONTEXT)) {
+    if (m.includes(prefix)) return limit
+  }
+  return 128000
+}
+
+export function contextUtilization(messages, model, configState = null) {
+  const tokens = estimateTokenCount(messages)
+  const limit = modelContextLimit(model, configState)
+  const ratio = limit > 0 ? Math.min(1, tokens / limit) : 0
+  return {
+    tokens,
+    limit,
+    ratio,
+    percent: Math.round(ratio * 100)
+  }
+}
+
+export function supportsNativeCompaction(providerType, model) {
+  if (providerType !== "anthropic") return false
+  const m = String(model || "").toLowerCase()
+  return m.includes("claude") && (m.includes("opus") || m.includes("sonnet"))
+}
+
+export function shouldCompact({ messages, model, thresholdMessages = DEFAULT_THRESHOLD_MESSAGES, thresholdRatio = DEFAULT_THRESHOLD_RATIO, configState = null, realTokenCount = null }) {
+  if (messages.length >= thresholdMessages) return true
+  const limit = modelContextLimit(model, configState)
+  const tokens = realTokenCount != null ? realTokenCount : estimateTokenCount(messages)
+  return tokens >= limit * thresholdRatio
+}
+
+export async function compactSession({
+  sessionId,
+  model,
+  providerType,
+  configState,
+  keepRecent = DEFAULT_KEEP_RECENT,
+  keepRecentTurns = DEFAULT_KEEP_RECENT_TURNS,
+  baseUrl = null,
+  apiKeyEnv = null
+}) {
+  const history = await getConversationHistory(sessionId, 9999)
+  if (history.length <= keepRecent + 2) return { compacted: false, reason: "too few messages" }
+
+  // Turn-based split: keep last keepRecentTurns complete turns
+  // A "turn" = one user interaction cycle (user msg + model response + all tool calls)
+  // Falls back to message-count if no turnId metadata is present
+  let splitIdx
+  const turnIds = []
+  const seenTurns = new Set()
+  for (const msg of history) {
+    if (msg.turnId && !seenTurns.has(msg.turnId)) {
+      seenTurns.add(msg.turnId)
+      turnIds.push(msg.turnId)
+    }
+  }
+  if (turnIds.length > keepRecentTurns) {
+    const keepFromTurnId = turnIds[turnIds.length - keepRecentTurns]
+    splitIdx = history.findIndex(msg => msg.turnId === keepFromTurnId)
+    if (splitIdx < 0) splitIdx = history.length - keepRecent
+  } else {
+    // Fallback: not enough turns, use message count
+    splitIdx = history.length - keepRecent
+  }
+  const toSummarize = history.slice(0, splitIdx)
+  const kept = history.slice(splitIdx)
+
+  // Layer 1: prune large tool outputs before sending to LLM
+  const pruned = pruneForSummary(toSummarize)
+  const summaryPrompt = pruned.map((m) => {
+    const content = m.content
+    if (Array.isArray(content)) {
+      return `[${m.role}]: ${content.map((b) => {
+        if (b.type === "text") return b.text || ""
+        if (b.type === "tool_use") return `[tool_use:${b.name}(${JSON.stringify(b.input || {}).slice(0, 120)})]`
+        if (b.type === "tool_result") return `[tool_result:${b.is_error ? "ERROR " : ""}${b.content || ""}]`
+        return ""
+      }).filter(Boolean).join("\n")}`
+    }
+    return `[${m.role}]: ${content}`
+  }).join("\n\n")
+
+  const hookPayload = await HookBus.sessionCompacting({
+    sessionId,
+    messageCount: history.length,
+    summarizeCount: toSummarize.length,
+    keepCount: kept.length
+  })
+  if (hookPayload?.skip) return { compacted: false, reason: "skipped by hook" }
+
+  let summaryText
+  let compactionUsage = null
+  try {
+    const response = await requestProvider({
+      configState,
+      providerType,
+      model,
+      system: COMPACTION_SYSTEM,
+      messages: [{ role: "user", content: summaryPrompt }],
+      tools: [],
+      baseUrl,
+      apiKeyEnv
+    })
+    summaryText = (response.text || "").trim()
+    compactionUsage = response.usage || null
+  } catch (error) {
+    return { compacted: false, reason: `compaction LLM call failed: ${error.message}` }
+  }
+
+  if (!summaryText) return { compacted: false, reason: "empty summary from LLM" }
+
+  // Replace all messages with: [summary] + [kept recent messages]
+  const summaryMessage = {
+    role: "user",
+    content: `<compaction-summary>\n${summaryText}\n</compaction-summary>`
+  }
+  await replaceMessages(sessionId, [summaryMessage, ...kept])
+
+  // Record compaction LLM usage so it's not "invisible"
+  if (compactionUsage) {
+    try {
+      const { pricing } = await loadPricing(configState)
+      const { amount } = calculateCost(pricing, model, compactionUsage)
+      await recordTurn({ sessionId, usage: compactionUsage, cost: amount })
+    } catch { /* best-effort */ }
+  }
+
+  await saveCheckpoint(sessionId, {
+    kind: "compaction",
+    iteration: 0,
+    compactedAt: Date.now(),
+    summarizeCount: toSummarize.length,
+    keepCount: kept.length,
+    summaryVersion: 1,
+    summaryLength: summaryText.length
+  })
+
+  return {
+    compacted: true,
+    summarizedCount: toSummarize.length,
+    keptCount: kept.length,
+    summaryLength: summaryText.length
+  }
+}